Subclonal response heterogeneity to define cancer organoid therapeutic sensitivity

Abstract

Tumor heterogeneity is predicted to confer inferior clinical outcomes with precision-based strategies, however, modeling heterogeneity in a manner that still represents the tumor of origin remains a formidable challenge. Sequencing technologies are limited in their ability to identify rare subclonal populations and predict response to treatments for patients. Patient-derived organotypic cultures have significantly improved the modeling of cancer biology by faithfully representing the molecular features of primary malignant tissues. Patient-derived cancer organoid (PCO) cultures contain subclonal populations with the potential to recapitulate heterogeneity, although treatment response assessments commonly ignore diversity in the molecular profile or treatment response. Here, we demonstrate the advantage of evaluating individual PCO heterogeneity to enhance the sensitivity of these assays for predicting clinical response. Additionally, organoid subcultures identify subclonal populations with altered treatment response. Finally, dose escalation studies of PCOs to targeted anti-EGFR therapy are utilized which reveal divergent pathway expression when compared to pretreatment cultures. Overall, these studies demonstrate the importance of population-based organoid response assessments, the use of PCOs to identify molecular heterogeneity not observed with bulk tumor sequencing, and PCO heterogeneity for understanding therapeutic resistance mechanisms.

Introduction

Cancer cell heterogeneity is a hallmark of human cancers and can arise from differences in the molecular profile or the metabolic states of cells. However this heterogeneity is not recapitulated in the common two-dimensional cancer cell cultures or syngeneic or transgenic murine models used for therapeutic development. Patient-derived cancer organoids (PCOs) have become the preferred in vitro culture technique for modeling tumor biology across many cancers^1,2,3,4,5. PCOs are generated with high efficiency, retain the molecular characteristics of their parent tissues, and can have responses that correlate with clinical outcomes for patients^6,7,8,9,10. To date, studies have examined the PCO response on the culture-well level, yet the heterogeneity between individual PCOs is not taken into consideration. Each well can contain hundreds of individual PCOs, each representing distinct clonal or oligoclonal units of the original tumor comprising 100–1000s of individual cells. Tools capable of capturing heterogeneity are of vital importance when subclonal heterogeneity can confer primary or acquired clinical resistance. As metabolic heterogeneity is a significant concern, we have developed the use of single-cell optical metabolic imaging (OMI) technologies and have adapted them for PCO response characterization across multiple cancer models^{8,13,14,15,14}.

Epidermal growth factor receptor (EGFR) inhibition, with cetuximab or panitumumab, is a standard targeted therapy option for patients with KRAS, NRAS, and BRAF^V600 wild-type metastatic colorectal cancer (CRC)^15,16,17. Many mechanisms of resistance to EGFR inhibition have been previously described, including metabolic changes and the development of subclonal heterogeneity^{18,19,20,21,22}. Acquired mechanisms of resistance following EGFR inhibition cannot be predicted for individual patients from pretreatment bulk sequencing or circulating tumor DNA (ctDNA) alone. The selective pressure of EGFR inhibition determines clones that either develop activating alterations in RAS signaling^23,24,25,26 or differential expression of key mediators across multiple pathways, including EGFR promoter methylation and amphiregulin-mediated signaling in the microenvironment^{27,28,29,30,31}. Clinical ctDNA can detect the acquisition of alterations associated with resistance to EGFR inhibition, but these most often are found at extremely low allele frequencies and cannot track important metabolic changes that occur in response to therapy³². The metabolic and molecular heterogeneity that develops in response to EGFR inhibition in CRC makes this an ideal clinical scenario to evaluate the use of PCOs to characterize this heterogeneity.

Here, we evaluate if assaying individual PCO responses, across a population of organoids, improves our ability to identify responsiveness to chemotherapy and radiation across a set of advanced cancers. Treatment response and organoid viability are determined using change in individual organoid diameter and optical metabolic imaging (OMI)^8,33. These imaging analyses determine the response to treatment across the population of PCOs, rather than current techniques of culture well-level assessments, to recognize subclonal populations that predict clinical response to therapy. We also integrate single-cell OMI analyses for assessing the metabolic response to chemotherapy. Next, to determine the potential to detect the acquisition of resistant subpopulations over time, CRC PCOs were treated with escalating doses of panitumumab, an EGFR inhibitor. Upon the development of resistance, each culture developed resistance by changing the expression of diverse pathways. Interestingly, we did not detect the acquisition of mutations in KRAS, BRAF, or EGFR. Overall, we demonstrate that PCOs can predict clinical activity of therapeutics across diverse treatments, identify clinically relevant heterogeneity using phenotypic and metabolic imaging, and identify PCO-line specific resistance mechanisms. These studies further validate organoid culture as a tool predictive of disease biology in translational studies.

Results

PCOs recapitulate driver mutations of parent cancers and commonly possess subclonal molecular heterogeneity

PCOs were developed from a dedicated tissue-procurement program from multiple cancer types and sample sources, including biopsy, fluid, and surgical specimens from patients (Suppl. Table 1). These samples were named using a code: locally advanced (L), metastatic (M), rectal (R), and colon (C). Other samples were obtained through the University of Wisconsin Carbone Cancer Center Precision Medicine Molecular Tumor Board referenced as “MTB”. To evaluate the mutation profile and potential for subclonal populations 33 unique PCO cultures were generated across cancers from 31 subjects. Across the PCO lines, 45 driver alterations were detected. There was 93% concordance for pairwise sets comparing next-generation sequencing (NGS) from clinical specimens and expanded PCOs (Fig. 1a). The minor discordance here is similar to the discordance seen when sampling a single patient’s cancer in separate regions^34,35. Subclonal mutations were identified in 91% of PCO cultures (Fig. 1b–f). PCOs expanded from two sites within the same tumor shared multiple clonal and subclonal alterations, however additional subclonal alterations could be expanded from a given site for LR9B (Fig. 1f).

Fig. 1

Subclonal molecular heterogeneity in PCOs. (a) Heatmap of pathologic alterations between pairwise tumor (black, T) and expanded PCOs (gray, O) for reported pathologic driver alterations with split designation (triangles) for multiple concurrent gene alterations. All variants colored on heatmap according to relative variant allele frequency (rVAF). All cases were mismatch repair proficient. (b) Analysis of subclonal non-synonymous variants (NSV) defined between 10–30% from cancer hotspot testing in PCOs. (c–e) Representative examples of clonality including clonal (black) and subclonal variance (gray) across representative cancers. (f) Multisite sampling from patient LR9 from endoscopic biopsy with VAF plotted including clonal (black) and discordant subclonal alterations in EXO1, MLH1, and KIT from LR9A (red) and LR9B (blue).

PCO growth and metabolic imaging independently track organoid heterogeneity

In addition to the above molecular heterogeneity, significant differences in the growth rate were observed within cultures (Fig. 2a). Furthermore, intra-organoid heterogeneity was seen by intrinsic auto-fluorescence of NAD(P)H and FAD lifetime and the optical redox ratio (ORR) observed across representative PCOs (Suppl. Figure 1). A total of 2096 individual PCOs were analyzed based on histology, stage, and disease status (Fig. 2b and Suppl. Table 1). There was a wide range of mean growth rates observed between PCO cultures (2–54%) and even wider range of growth rates of individual organoids within the cultures (lowest range + 20.6% to highest range of + 201.7% (Fig. 2c).

Fig. 2

PCO heterogeneity in growth. (a) Representative PCO brightfield microscopy (patient MC1, see Supp. Table 2) annotated with percent difference in individual organoid diameter. (b) Baseline clinical characteristics for cultures colored by the key below, including pathologic review of tumor grade and clinical parameters of stage from tumor (T), node (N) and metastases (M) combined to complete diagnostic staging (Stage). Current disease status noted (Recurrence in black). (c) Dot plots colored by the key below for primary tumor type (Primary) for localized rectal (light green), metastatic rectal (dark green), localized colon (light blue), metastatic colon (dark blue), breast (purple), ovarian (beige), gastrointestinal stromal tumor (GIST, gray), and lung (white) with each point representing an individual organoid for normalized Δ diameter at 48 h, median denoted with vertical bar (orange).

Previously, we determined that individual baseline organoid culture conditions are not responsible for the heterogeneity observed in both murine and patient-derived organotypic cultures³⁶. Baseline diameter, relative passage number, density per field of view, and distance from the edge of the matrix were not predictive of organoid growth³⁶. CRC PCOs were analyzed for individual organoid response to chemotherapy with 5-FU and oxaliplatin (FOLFOX) yielding differences in individual organoid response in tumor growth (Fig. 3a) and single cell metabolism (Fig. 3b). To make certain that this response heterogeneity was not secondary to the culture technique and drug penetration in to the matrix, the location of individual organoids within the 3D matrix was evaluated and not found to predict therapeutic response to FOLFOX or panitumumab, a monoclonal antibody against EGFR (Fig. 3c and Suppl. Figure 2a,b). Additionally, across five independent CRC PCO cultures, baseline growth rate had limited correlation with treatment response (Suppl. Figure 3).

Fig. 3

FOLFOX resistance monitoring reveals subclonal resistance from a patient with familial adenomatous polyposis. (a) Representative therapeutic study (patient LC1) with FOLFOX imaged with brightfield microscopy at 0 h and 48 h annotated with percent difference in individual organoid diameter and (b) representative images of organoids with two-photon ORR at 48 h. (c) Comparison of PCO location perpendicular to the matrix edge correlated against growth with FOLFOX treatment across three independent cultures (LR4, LR5, MC7) along with the coefficient of determinant (R²). (d) Sites of tissue sampling from patient with multiple site polyp (n = 4) and tumor sampling (n = 5). (e) Heatmap of pathologic alterations of PCOs derived from individual polyps (P1–P4) and tumors (T1–T5) compared between primary tissue (black, T) and organoids (gray, O) plotted as relative variant allele frequency (rVAF). Denoted are tumor suppressor genes (green) and oncogenes (red) plotted as a function of rVAR. (f) Heatmap of expanded PCO subclones selected by individual spikes using NGS profiling. (g) Comparison of normalized Δ diameter and FLIRR at 48 h stratified by parent culture, FBXW7^WT (wild type, WT), and FBXW7^R479Q (mutant, MT) using two-sided student t-test (p > 0.05). (h) Representative PCOs metabolism assessed at 48 h by ORR (NAD(P)H/FAD) for control (top panel) and FOLFOX stratified by FBXW7 profile. Scale bar represents 100 µm. (i) Heatmap of OMI parameters by FBXW7 status stratified with respective Z-score as compared to parent culture. Significance noted for |GΔ| > 0.75 for individual OMI parameters with corresponding positive (black *) or negative (white *) effect size. Z-score defined by ${\overline{\text{x}}}_{{{\text{PCO}}}}$ (average value of individual OMI parameter for individual PCO culture), ${\overline{\text{x}}}_{{{\text{population}}}}$ (average value of individual OMI parameter across the control population), σ_population (standard deviation of an individual OMI parameter across the control population) for the control conditions ${\overline{\text{x}}}_{{{\text{population}}}}$ and σ_population refer to parent culture values. (j,k) Gaussian distribution plots of normalized PCO diameter change assessed from 0 to 48 h including control (gray), 5-FU (blue), oxaliplatin (red), and FOLFOX (purple). Molecular profile at FBXW7 denoted wildtype (WT, solid line) and mutant FBXW7^R479Q (MT, dashed line). Response assessed using effect size (GΔ) relative to untreated control stratified by molecular profile at FBXW7 for (j) normalized Δ diameter and (k) ORR at 48 h. Scale bars for brightfield (black bar) represent 200 µm, scale bars for OMI (white bar) represent 100 μm.

Subculturing of PCOs can detect organoids with rare driver alterations that have potential to alter therapeutic sensitivity

A patient with familial adenomatous polyposis syndrome with a germline pathologic alteration in the Adenomatous Polyposis Coli gene (APC^Q789*) had PCOs derived (LC1) by multi-site sampling of 4 polyps and 5 distinct regions of a single transverse colon cancer (Fig. 3d). Individual organoid subculturing experiments were performed to evaluate the potential to identify subclones with distinct mutation profiles and to evaluate the potential subclones with molecular changes to have differential sensitivity to standard chemotherapy. Pathologic alterations in APC were present in tissue from all polyp and cancer samples. Changes in normalized variant allele frequency (nVAF) were observed between patient tissue and the expanded PCOs, respectively (APC^R499* in polyp 1, nVAF 41% v. 89%; APC^R554* in polyp 4, nVAF 28% v. 49%; APC^R1450* in tumor 2, nVAF 0% v. 50%; and BRAF^{V600_K601delinsE} in Tumor 3, nVAF 0% v. 48%; Fig. 3e). To determine if subclonal populations could be further identified, individual PCOs were selected for further subculture expansion (Fig. 3f). Three independent PCO cultures from individual organoids from polyp 1 possessed the APC^R499* mutation. Expanded cultures from polyp 2 harbored diversity in APC mutations, including the known germline APC^Q789*, as well as APC^R499* which was not detected in the parent culture in the initial analysis. Re-analysis revealed the presence of the APC^R499* mutation in the parent culture at a frequency just under the detection cutoff for VAF analysis (cutoff 10%; mutation detected in 97/1025 reads; 9.4%). Interestingly, for all subculture expansions of polyp 2, only one or the other mutation was detected to any significant level (Subculture 1 APC^R499* mutation detected in 3/816 reads; Subculture 2 APC^Q789* mutation detected in 2/510 reads; Subculture 3 APC^Q789* mutation detected in 0/271 reads). Spikes from tumor 2 harbored FBXW7^R479Q (nVAF 49%) which was not detected by sequencing of the primary tissue or from either the parent organoid culture or two of the four expanded PCO spikes.

Next, the treatment response to standard FOLFOX chemotherapy was compared between those organoids with wildtype (WT) or mutant FBXW7^R479Q (MT) to characterize the potential for differential sensitivity related to this pathologic alteration. Fluorescence Lifetime Imaging Redox Ratio (FLIRR) enables metabolic comparison of organoids captured on different days as it overcomes several experimental limitations of fluorescence intensity imaging. Fluorescence intensity measurements are influenced by laser power, detector gain, light scattering, and the concentration of the fluorophore; however, fluorescence lifetimes are independent of these experimental factors^37,38. Although the FLIRR is a useful measurement of cellular metabolism, it is not a substitute for the optical redox ratio and instead represents an additional variable to compare metabolic endpoints. Relative to the parent population, the growth rate and FLIRR were unchanged in the FBXW7^WT and FBXW7^R479Q PCOs (Fig. 3g). There was not a detectable difference in the ORR between these PCO lines prior to treatment (Fig. 3h). The ORR was next evaluated in response to chemotherapy with FOLFOX. For FBXW7^WT PCOs the ORR was reduced after FOLFOX treatment indicating a robust response, in contrast to the FBXW7^R479Q PCOs (Fig. 3h). Further analysis revealed changes for NAD(P)H lifetime parameters τ_m, τ₁, and τ₂ with FOLFOX treatment in FBXW7^WT PDCOs, and to a lesser degree in FBXW7^R479Q PCOs (Fig. 3i). Response as assessed by normalized change in diameter was similar for 5-FU alone (GΔ: MT 1.6 versus (v) WT 1.4). FBXW7^R479Q conferred reduced sensitivity for oxaliplatin (GΔ: MT 2.2 v WT 3.0), and FOLFOX (GΔ: MT 2.3 v WT 3.0; Fig. 3j). Consistent with changes in PCO diameter, metabolic response using ORR was reduced in FBXW7^R479Q compared to FBXW7^WT PCOs for 5-FU (GΔ: MT 0.6 v WT 1.5), oxaliplatin (GΔ: MT 1.9 v WT 3.2), and FOLFOX (GΔ: MT 2.3 v WT 3.1; Fig. 3k).

Organoid-level growth and single-cell OMI identify EGFR inhibition sensitivity across populations of CRC PCOs

It is well-established in metastatic CRC that KRAS, NRAS, and BRAF^V600 mutations lead to clinical therapeutic resistance to the EGFR monoclonal antibodies, panitumumab and cetuximab³⁹. The activity of panitumumab was assessed in a panel of CRC PCOs stratified by these resistance mechanisms (Fig. 4). Population modeling after EGFR inhibition in MC4, possessing a RAS^A146V mutation, revealed two distinct populations both with persistent growth. In contrast, MR3 (RAS^WT/RAF^WT) PCOs converged to a single population that achieved growth arrest, with most organoids decreasing in size after EGFR inhibition (Fig. 4a,b). The panitumumab sensitivity was assayed across a panel of RAS^MT or RAF^MT PCOs (n = 4) and RAS^WT/RAF^WT PCOs (n = 9). In a pooled analysis, RAS^MT or RAF^MT PCOs had minimal changes in size between treatment (GΔ: 0.51; Fig. 4c). In the pooled analysis of RAS^WT/RAF^WT PCOs, measurements at 48 h alone did not change significantly with treatment. However, the normalized change in diameter with individual organoid tracking across the population did identify a large effect size (GΔ: 1.3; Fig. 4d), indicating the enhanced sensitivity of assessing treatment response on the individual organoid level over a treatment course. These results were consistent with individual culture responses across the PCO lines, with significant differences in effect sizes observed depending on the RAS/RAF status (p < 0.005; Fig. 4e). No PCOs with RAS/RAF mutations achieved an effect size ≥ 0.75, whereas the RAS/RAF WT PCOs had a range of responses by effect sizes (GΔ 0.74–2.1). Clinically, the patient’s cancer from which MR3 was derived was controlled with panitumumab at the site of the biopsy, which was used for PCO generation. However, the patient had disease progression at distant sites with the acquisition of KRAS amplification (copy number = 89; Fig. 4f).

Fig. 4

Assessment of PCO response to EGFR inhibition. (a) Representative brightfield images of therapeutic resistance of KRAS^A146V MC4 with persistent growth of control and panitumumab from 0 to 48 h in contrast to (b) RAS^WT MR3 with growth arrest assessed by normalized Δ diameter at the organoid level. (c) Pooled analysis of diameter for four independent lines predicted for resistance to EGFRi: RAS^MT (LR2, MR11, MC4) and BRAF^V600E (MC5A) at 48 h (left panel) and change in diameter at 48 h (right panel) by assessment of individual PCOs normalized to baseline diameter at 0 h with corresponding effect size across distributions (GΔ). (d) Pooled analysis of diameter for nine independent RAS^WT/BRAF^WT PCOs at 48 h (left panel) and change in diameter at 48 h (right panel) by assessment of individual PCOs normalized to baseline diameter at 0 h with corresponding effect size across distributions (GΔ). (e) Line-specific sensitivity plotted by effect size (GΔ) including RAS^WT/BRAF^WT (gray) compared against RAS^MT (red) and BRAF^V600E (violet) using student’s t-test for effect size of normalized Δ diameter. (f) MR3 PCO response of single agent EGFR inhbition (panitumumab) *denoted in (e) represents a prospective clinical assessment tracked on CT scan at a 15-week follow-up showing local disease control at the biopsy site and a non-target progression in the right upper lung. Green lines indicate longest diameter (LD) of adrenal metastasis at baseline and restaging and measurements of non-target disease progression in right lung.

PCOs response predicts patient’s clinical treatment response

Clinical response was prospectively correlated from PCOs developed from 13 patients with advanced cancers and 4 controls with canonical resistance mutations for EGFR inhibition (Suppl. Table 2). The effect of single agent gemcitabine on MTB-3 (ovarian cancer) was negligible (GΔ 0.40; Fig. 5a,b) consistent with a clinical outcome of progression in retroperitoneal adenopathy (Fig. 5c). However, MTB-3 had intermediate sensitivity to single agent paclitaxel (GΔ 1.1, Fig. 5b). Metastatic CRC MC7 was collected at diagnosis and PCOs demonstrated significant response to FOLFOX (GΔ 2.2, Fig. 5d,e), which matched the durable clinical response in the liver (Fig. 5f).

Fig. 5

Validation of PCO response for clinical prediction. (a) Representative brightfield microscopy from MTB-3 ovarian (Ov) PCOs at baseline and 48 h; scale bar represents 200 μm for each panel. (b) Gaussian distributions for growth at 48 h with respective effect sizes (GΔ) for MTB-3 Ov PCOs treated with gemcitabine 50 μm (24 h, green), paclitaxel 50 nM (48 h, gold) or control (black) as assessed at 48 h. (c) Clinical response from the initial restaging CT scan of subject MTB-3 confirming the disease with enlarging retroperitoneal adenopathy after treatment with single agent gemcitabine on CT imaging. (d) Representative brightfield microscopy for MC7 PCOs treated with control (top panels), or FOLFOX (5-FU 10 μm and oxaliplatin 5 mμ (e) Gaussian distributions of MC7 for Δ diameter over 48 h and respective effect sizes (GΔ) for 5-FU 10 μm (blue), oxaliplatin 5 μm (red), and FOLFOX (violet). (f) Restaging CT scan of MC7 shows partial response after FOLFOX. (g) Experimental sensitivity with clinical outcome or canonical mechanism of resistance labeled by treatment type including chemotherapy (purple), targeted therapy (blue), canonical EGFRi resistance (RAS^MT or RAF^MT, red), and radiation (black) with reported significance by two-sided student t-test. (h,i) Comparison of FOLFOX effect size (GΔ) between PCOs with disease progression after FOLFOX chemotherapy versus subjects without prior drug exposure assessed using two-sided student t-test with prior established sensitivity thresholds (shaded region). (h) Absolute diameter effect size assessed at 48 h for single agent 5-FU (ns), oxaliplatin (ns) and FOLFOX (*p < 0.05) between clinically resistant and unknown cohorts. (i Effect size of growth (percent Δ diameter) tracked from 0 to 48 h for single agent 5-FU (*p < 0.05), oxaliplatin (**p < 0.005) and FOLFOX (***p < 0.0005) between clinically resistant and unknown cohorts. (j) Bar plot of negative predictive value (NPV) and positive predictive value (PPV) for prospectively treated subjects. (k) Receiver operator curve (ROC) in response prediction plotted as false positive rate versus sensitivity with the colored line showing the continuum of effect size (GΔ) for change in diameter and corresponding area under the curve (AUC).

PCO response assessment was expanded across a diverse set of advanced solid cancers. Clinical outcomes were compared by radiographic imaging or canonical mechanism of therapeutic resistance to EGFR inhibition (Fig. 5g). To identify thresholds for therapeutic resistance, we used PCOs with clinical resistance or otherwise unknown clinical response (Fig. 5h,i). As part of this analysis, we evaluated the organoid response using 2 different methods. In the first analysis, after 48 h of treatment the diameter of residual organoids was assessed relative to control and conferred limited sensitivity in predicting clinical sensitivity (Fig. 5h). In contrast, the normalized change in diameter using individual organoid tracking improved sensitivity for identifying response to 5-FU (p < 0.05), oxaliplatin (p < 0.005), and FOLFOX (p < 0.001; Fig. 5i). For those patients whose PCOs had a GΔ > 1.25 in response to the treatment that they received clinically, a partial response (at least 30% reduction in cancer diameter per RECIST v1.1) was observed in all cases (Fig. 5j). Additionally, if the patient’s PCOs were limited to a GΔ < 0.75, no partial responses to clinical treatment were observed (Fig. 5j). Half of those patients whose PCOs had a GΔ between 0.75 and 1.25 achieved a partial response (Fig. 5j). A receiver operator curve for partial response prediction achieved AUC of 0.987 as based on GΔ achieved (Fig. 5k).

Longitudinal evaluation of PCO population response to dose escalation of EGFR inhibitors to measure therapeutic efficacy

As the PCO growth evaluations do not require labels, dyes, or organoid destruction, these methods lend themselves to longitudinal assessments. Growth rate was captured by normalized change in diameter over 96 h. Dosing was started at 20% of the panitumumab dose able to be achieved in the blood of patients treated clinically. If a given culture was found to have a normalized change in diameter of > 20%, the culture was escalated to the subsequent dose level with a 20% increase in panitumumab concentration. The analysis was performed to facilitate adaptive dosing weekly based on the 96 h growth indices. The longitudinal tracking of resistance was compared using time to resistance (TTR), defined as the time to persistent growth at 100% physiologic C_max of panitumumab.

To improve the accessibility of OMI methods epifluorescence wide-field microscopy was used to assay panitumumab sensitivity during dose escalation (Suppl. Figure 4). As a representative example, the mean values of ORR for MC7 were relatively stable as the PCOs continued to grow with dose escalation (Suppl. Figure 4b). When visualized as normalized Gaussian distributions, divergent populations developed at week 3 (Suppl. Figure 4c). The culture continued to achieve persistent growth at 100% physiologic C_max panitumumab which shifted to an increase in the ORR by week 6 of dose escalation (p < 0.001; Suppl. Figure 4b-c).

RAS MT PCOs achieved serial thresholds for growth each week with ex vivo resistance determined within 30 days (Fig. 6a and Suppl. Figure 4d). Across the RAS/RAF WT PCOs a wide range of TTR was observed (Fig. 6b and Suppl. Figure 4e). Primary resistance was identified in RAS WT cultures (MR9, MC7, LR6, and LR4) with a TTR of < 40 days. Additional lines had intermediate TTR (LR5, MR8) between 40–100 days, and prolonged sensitivity either achieving therapeutic resistance in over 100 days (MC1) or had complete arrest of proliferation due to therapeutic sensitivity (MR3; Fig. 6b, and Suppl. Figure 4e).

Fig. 6

Generation and molecular profiling of CRC PCOs with ex vivo resistance to EGFRi. (a,b) Time course of dose escalation stratified by RAS mutation profile including MT and WT. (c) Serial molecular profiling by cancer hotspot next generation sequencing of PCOs over the course of dose escalation with heat map labeling of absolute passage (P_x, blue-yellow), physiologic C_max of EGFRi (black-red), and alteration variant allele frequency (VAF, black-green) with split cells (triangle) representing multiple gene-specific alterations. (d) Upset plot for individual genes from RNASeq transcriptional profiles compared in triplicate between control and resistance to EGFRi panitumumab. Shown are clinical molecular profiles obtained at time of PCO profiling including designation for KRAS amplification (red +) and nVAF in PIK3CA (green), PTEN (green), all with concurrent alterations in TP53 (light blue). Size of intersection shown colored by comparison for unique genes for a single PCO (white), two PCOs (light gray), three PCOs (dark gray) or shared between 4 + lines (black) ordered in descending order. Black circles denote comparisons with individual gene alterations plotted for individual lines. Significance defined by p_adj < 0.05. (e) Upset plot for GSVAs plotted against molecular profile as outlined in (d) with significance defined by p_adj < 0.1 accounting for sign change using -log₁₀(p_adj)*sign(FC). FC, fold change.

Diverse molecular heterogeneity at ex vivo resistance to EGFR inhibition within and across PCO lines

To further evaluate the mechanisms by which panitumumab conferred resistance, pairwise NGS using panel-based DNA profiling was performed comparing matched baseline and panitumumab resistant PCOs after dose escalation. Repeat DNA sequencing did not reveal the acquisition of pathologic alterations in RAS, BRAF, or extracellular kinase alterations in EGFR, which are commonly identified clinically, across eight independent RAS/RAF WT PCOs once resistant to panitumumab (Fig. 6c).

To further characterize the molecular heterogeneity at the level of transcription, RNAseq was performed in triplicate between baseline and EGFR inhibitor resistant RAS/RAF WT PCOs. Upset plots were generated to compare shared genes and pathways across resistant profiles from nine unique RAS/RAF WT PCOs at panitumumab resistance (Fig. 6d,e). Using a negative binomial generalized linear model, TTR alone was not predictive of significant differences in the numbers of genes (p = 0.20) or pathways (p = 0.55). PCOs with pre-existing alterations that could lead to relative panitumumab resistance (ie. KRAS amplification, PTEN loss, PIK3CA alterations) were found to have fewer genes with significant differential expression at EGFR inhibitor resistance (predicted using a negative binomial generalized linear model (p < 0.002; Fig. 6d and Supplemental Materials). Similar findings were seen in gene set pathway enrichment analysis with a divergence of statistically altered pathways for cultures with relative panitumumab resistance (p = 0.002; Fig. 6e).

Many thousand individual transcripts were found to have statistical differences in pairwise sets across the nine RAS/RAF WT cultures at resistance to EGFR inhibitor (defined with p_adj < 0.05). However, individual transcripts for a given culture were highly variable between lines at panitumumab resistance (Fig. 7a). The downregulation of TIMP1 was the only transcriptional change that was statistically concordant between all nine resistant cultures. Gene set variation analysis revealed heterogeneity in pathways significantly altered at resistance to EGFR inhibition (Fig. 7b and Suppl. Figure 5). The diversity of pathways at panitumumab resistance in CRC PCOs included TGFβ signaling, WNT signaling, cell adhesion molecule expression, MAP kinase signaling, with rare events noted in homologous recombination, mismatch repair, and glutathione metabolism when mapped against KEGG (Fig. 7b) and Hallmark gene set pathways (Suppl. Figure 6). No pathways were uniformly altered across all the PCO lines at panitumumab resistance.

Fig. 7

Heterogeneity in transcriptional resistance to EGFR inhibition in RAS^WT/RAF^WT CRC PCOs. (a) Heatmap for pairwise comparison from base culture to resistance under selection of C_max panitumumab with differential expression (n = 3) with significance shown with black box defining differential expression for individual genes if p_adj < 0.05 plotted as function of log₂(FC). (b) Circa plot of KEGG pathway enrichment in RAS^WT/RAF^WT CRC organoids. Each contiguous arc represents individual PCO lines with differential expression at EGFR inhibitor resistance organized by pathways grouped by individual colors including TGFβ (red), WNT signaling (orange), cell adhesion molecules (yellow), MAP kinase signaling (green), homologous recombination (blue), mismatch repair (violet), and glutathione metabolism (brown). Significance shown with black box defining differential expression for individual genes if p_adj < 0.05 and plotted as function of log₂(FC).

Discussion

There remains an urgent clinical need to develop personalized therapeutic strategies for patients with cancer. PCOs remain an exciting tool for understanding cancer biology and fostering drug development. This technology provides a potential mechanism for prediction of clinical response^8,9,10. As more laboratories are expanding their use of PCOs for translational studies, detailed investigations are needed to identify the limitations of these models and how to best take advantage of the numerous strengths, including the representation of cancer cell heterogeneity for individual patients. Here we demonstrate PCO heterogeneity using a dedicated assessment of individual organoids tracked over time.

Taking the heterogeneity into consideration, we demonstrate the importance of therapeutic assessments using population modeling based on individual organoid tracking over time and single cell-level metabolic imaging. We show the ability to expand individual organoids demonstrating that distinct populations exist that can confer differential response to chemotherapy. Individual organoids response provides an alternative to ATP-based viability assays that are read without taking individual organoid response into consideration^8,19,36. The assessment of organoid growth as compared to organoids size was shown to improve the sensitivity response using anti-EGFR inhibition.

Longitudinal time-course modeling in organoids provides many advantages. This technique was independently developed to overcome the lack of sensitivity to EGFR monoclonal antibodies reported in prior ATP-based viability assays¹⁹. TTR across these samples was of clinical relevance with serial resistance in all RAS^MT cultures, yet a range of sensitivities was observed in the RAS^WT cultures in a clinically relevant range of weeks to few months. Longitudinal evaluation of therapeutics using PCOs needs further clinical validation as a marker of duration of clinical response. Additionally, further studies are needed to determine if the mechanisms of resistance identified in PCOs also represent the mechanisms that occur clinically in the cancer from which the organoids are derived. It is of particular interest that the RAS^WT/RAF^WT CRC PCOs were able to become resistant to anti-EGFR therapy in relatively short intervals with each of the investigated PCO lines becoming resistant. Impressively, the PCO RNA sequencing highlights the complexity of resistance mechanisms identified within each line and the heterogeneity of mechanism observed between lines, with no pathways uniformly altered across the PCO lines. There is a critical need in the field of precision oncology to identify patient specific resistance mechanisms, however, to date these assays have not been performed in a clinically relevant time frame. Of note, across these PCO lines investigated, we did not identify the acquisition of resistance mutations within our organoid cultures. Using circulating tumor DNA (ctDNA), it is quite common to identify the acquisition of mutations or other alterations in RAS, RAF, or EGFR, among others, though these are commonly at exceedingly low variant allele frequencies. The PCO data here calls to question the importance of these rare alterations in clinical resistance and could lead to future combined analyses with PCOs and ctDNA offering complementary data in identifying resistance mechanisms for individual patients.

This study details PCO response assessment to define the contributions of subclonal populations for clinical response prediction. The contributions of experimental parameters including organoid size, passage, and plating density did not predict differential growth across populations. Rather, it was line specific heterogeneity that could be readily selected using the expansion of individual PCOs, that persisted under therapeutic treatment, and that drove a prediction of clinical outcomes. Despite the promise of defining these unique populations, the complexities of transcriptional regulation remain formidable with resistance including EGFR inhibition after dose escalation. This work provides a framework to characterize line specific adaptive mutability in response to targeted therapy⁴⁰ and has been integrated into the prospective investigation of EGFR inhibition for advanced CRC (NCT04587128).

Methods

Cell isolation and organotypic culture techniques

All studies were completed following University of Wisconsin Health Sciences Institutional Review Board (IRB) approval with informed consent obtained from subjects through the University of Wisconsin (UW) Molecular Tumor Board Registry (UW IRB#2014-1370) or UW Translational Science BioCore (UW IRB#2016-0934). Research herein conforms to the principles outlined in the Declaration of Helsinki. Briefly, tissue obtained from needle, endoscopic biopsy, or primary surgical resection was placed in chelation buffer for at least 30 min. Two PBS washings were performed for tissues acquired using endoscopic biopsies. Digestion was performed in DMEM stock (Supplementary Table 3) with the addition of collagenase (50 mg/mL) and dispase (10 mg/mL). Tissue was digested with intermittent shaking and mechanical disruption using a p1000 pipette from 15 min to 2 h dependent upon tissue size. Malignant fluids (ascites, pleural effusion, or pericardial effusions) were initially pelleted and separated using a Ficoll solution preparation (Sigma). Once digestion or Ficoll preparation was completed, the samples were centrifuged at 1000 rpm at 4 °C for 5 min and resuspended in ADF stock (Supplementary Table 3). PCO suspensions were immediately mixed at a 1:1 ratio with Matrigel matrix (Corning). Droplet suspensions were plated and set for three to five minutes at 37 °C then inverted for at least 30 min to solidify the matrix and avoid the direct interface of PCOs with the plastic interface. Plated cultures were overlaid with 450 µL of feeding medium supplemented with tissue dependent components (Supplementary Table 3) and incubated at 37 °C in 5% CO₂. Medium was replaced every 48–72 h.

Therapeutics studies

PCOs were collected from 24-well culture plates and passaged to single well 35 mm glass dishes (Cellvis) or 24-well glass plates. Images were taken on a Nikon Ti-S inverted microscope using a 4 × objective prior to treatment. Pharmacologic agents were prepared at physiologic C_max including 5-fluorouracil (5-FU) [10 µM]⁴¹, oxaliplatin [5 µM]⁴², SN-38 [1.5 nM]⁴³, gemcitabine [50 µM]⁴⁴, paclitaxel [50 nM]⁴⁴, olaparib [200 nM]⁴⁵, osimertinib [30 nM]⁴⁶, and panitumumab [1.6 µM]⁴⁷ (Supplementary Table 4). The duration of therapy was continuous over 48 h for all agents with the exception of gemcitabine which was dosed over 24 h to model clinical pharmacokinetics⁴¹. Given the uncertainty of monoclonal antibody (mAb) delivery at the time of study design, EGFR inhibition assays were performed after 1-h of preincubation of the PCOs that were suspended in physiologic panitumumab (Supplementary Fig. 7). Pharmacologic agents were received from the University of Wisconsin Carbone Cancer Center Pharmacy including 5-FU (Fresenius), gemcitabine (Hospira), oxaliplatin (Hospira), paclitaxel (Athenex), and panitumumab (Amgen). Additional agents prepared from stock powders including SN-38 (Sigma), olaparib (LC Laboratories), and osimertinib (LC Laboratories). DMSO concentration was < 0.1% v/v for all final culture conditions. Radiation was delivered using XStrahl RS-225 cabinet delivering a dose rate of approximately 3.27 Gy/minute using an aluminum filter with monthly quality assurance performed by the UW Medical Radiation Research Calibration Lab. After radiation, media was exchanged at 48 h and response relative to control was assayed at 96 h. All therapeutic studies were performed as three independent biological replicates.

EGFR inhibition with dose escalation studies

Following maturation, PCOs were treated with stepwise dose escalation. Cultures were started at 20% physiologic C_max panitumumab (46ug/ml) and assessed for growth (normalized change in diameter at 96 h). If cultures surpassed ≥ 20% growth, stepwise dose escalation was repeated with an interval increase of 20% C_max. Time to resistance (TTR) was defined by duration of time from the initiation of dose escalation to persistent growth at 100% C_max panitumumab. Epifluorescence signal was collected using a Nikon Ti-S epifluorescence microscope equipped with a 4X air objective (Nikon CFI Plan Fluor, NA 0.13, FOV: 5.5 mm × 5.5 mm) and CMOS camera (Flash4, Hamamatsu). NAD(P)H fluorescence was excited with a DAPI filter cube (Nikon, ex: 361-389 nm / em: 435-485 nm, integration time 100 ms) and FAD fluorescence was excited by Intensilight C-HGFI (Nikon) through a custom filter set (Semrock, ex: 426-486 nm / em: 525.5–630.5 nm, integration time 200 ms). Epifluorescence signal was thresholded in ImageJ based on NAD(P)H signal intensity to define organoid-level regions of interest (ROI). ROIs were used to measure the mean intensity for the optical redox ratio, defined as the intensity of NAD(P)H/FAD.

Brightfield and epifluorescence image analysis

Normalized change in diameter as an assessment of growth was measured from brightfield images collected at 0 h and 48 h using ImageJ v1.51–1.54 (https://imagej.net/ij/) and converted from pixel to distance using the triplicate measurements of a 1000 μm standard. The distance to the edge of the Matrigel was calculated from the closest edge of the PCO with a perpendicular axis to the matrix edge to represent the shortest length. Experimental correlations between baseline size, relative passage number and PCOs per field of view were compared to growth using the adjusted coefficient of determination (R²) by OriginPro 2020 (https://www.originlab.com/). Effect size was determined using Glass’s Delta (GΔ) which normalizes the difference in mean observed signal to the control population normalized to the standard deviation of the control⁴⁸. Interquartile ranges were compared using effect size normalized to the first quartile. Epifluorescence signal was thresholded in ImageJ to NAD(P)H signal intensity to define organoid-level regions of interest (ROIs). The ROIs were then used to measure the mean intensity for the organoid-level optical redox ratio (ORR) defined as intensity ratio of NAD(P)H/FAD.

Two-photon optical metabolic imaging (OMI) acquisition and analysis

Briefly, NAD(P)H and FAD were excited at 750 nm and 890 nm, respectively, using a tunable Ti:sapphire laser (Coherent, Inc), an inverted microscope (Nikon, Eclipse Ti), and a 40 × water immersion (1.15NA, Nikon) objective. Both NAD(P)H and FAD images were obtained for the same field of view. FAD fluorescence was isolated using an emission bandpass filter of 550/100 nm, while NAD(P)H fluorescence was isolated using an emission bandpass filter of 440/80 nm. Fluorescence lifetime data was collected using time-correlated single-photon counting electronics (SPC-150, Becker and Hickl) and a GaAsP photomultiplier tube (H7422P-40, Hamamatsu). Images (512 × 512 pixels) were obtained using a pixel dwell time of 4.8 μs over 60 s total integration time. A Fluoresbrite YG microsphere (Polysciences Inc.) was imaged as a daily standard for fluorescence lifetime. The lifetime decay curves were fit to a single exponential decay, and the fluorescence lifetime was measured to be 2.1 ns (n = 7), which is consistent with published values^11,49,50.

Fluorescence lifetimes were extracted using SPCImage software (SPCImage v8.1, Becker & Hickl, https://www.becker-hickl.com/products/spcimage/). A bin of 3 × 3 pixels was used to maintain spatial resolution, the fluorescence lifetime decay curve was convolved with the instrument response function and fit to a two-component exponential decay model using,

$$I\left(t\right)= {\alpha }_{1}exp\left(-\frac{t}{{\tau }_{1}}\right)+{\alpha }_{2}exp\left(-\frac{t}{{\tau }_{2}}\right)+C$$

where $I\left(t\right)$ is the fluorescence intensity at time $t$, $\alpha$ is the fractional contribution of each component, $\tau$ is the lifetime of each component, and C accounts for background light. The two lifetime components are used to distinguish between the free and bound forms of NAD(P)H and FAD^51,52. The mean fluorescence lifetime was calculated using, ${\tau }_{m}={\alpha }_{1}{\tau }_{1}+{\alpha }_{2}{\tau }_{2}$. The decay curves for NAD(P)H and FAD were integrated for each pixel to obtain intensity values. The optical redox ratio (ORR) was calculated by dividing the intensity of NAD(P)H by the intensity of FAD. All reported ORRs are normalized to the average control ORR of the same patient and time point.

For the analysis of FLIRR, a semi-automated cell segmentation algorithm was developed using Cell Profiler software (v4.0.7, https://cellprofiler.org/previous-releases) ⁵³. This system identified pixels belonging to nuclear regions using a customized threshold code. Cells were recognized by propagating out from the nuclei within the image. To refine the propagation and to prevent it from continuing into background pixels, an Otsu Global threshold was used. The cell cytoplasm was defined as the cell borders minus the nucleus. Values for NAD(P)H lifetime variables (τ_m, τ_1, τ_2, α_1, α_2), FAD lifetime variables, NAD(P)H intensity, and FAD intensity were averaged for all pixels within the cytoplasm of each cell, as described previously^54,55,56. FLIRR was calculated by dividing the fractional contribution of bound NAD(P)H (α2) by the fractional component of bound FAD (α1)^37,38,57,58

Subpopulation modeling to characterize response heterogeneity

The collective cell population of each treatment type was input into a Gaussian mixture distribution model⁵⁹ (MATLAB, version 2018a, MathWorks, Natick, Mass, https://www.mathworks.com/products/matlab.html) given by:

$$f\left(y;{\Phi }_{g}\right)=\sum_{i=1}^{g}{\pi }_{i}\phi (y;{\mu }_{i},{V}_{i})$$

where $g$ is the number of subpopulations, $\phi \left(y;{\mu }_{i},{V}_{i}\right)$ is the normal probability density function with mean ${\mu }_{i}$, variance ${V}_{i}$, and ${\pi }_{i}$ is the mixing proportion. Goodness of fit was calculated given a set of subpopulations ($g$ = 1, 2, or 3) using an Akaike information criterion⁶⁰. The number of subpopulations was determined based on the lowest Akaike score. Probability density functions were normalized to ensure that the area under the curve for each treatment group was equal to 1. Treatment effect size was calculated using GΔ⁴⁸ and defined as meaningful if > 0.5 for single cell OMI and > 1.5 for sphere growth analysis based on prior work⁸. Due to expected heterogeneity across mixed populations, the pooled analyses of multiple cultures in response to EGFR inhibition was performed by Gaussian population modeling using GraphPad Prism 9 (GraphPad Software, Boston, MA, USA, www.graphpad.com).

Next generation sequencing of PCOs for DNA and RNA

PCOs were collected from ~ 4–6 wells, washed in 1 × sterile PBS, and stored at − 80 °C as a cell pellet until DNA and RNA isolation. For RNA collection, following washing, PCOs were stored at − 80 °C in RNA_later (Qiagen) until library preparation. DNA was isolated on a Maxwell 16 AS2000 or Maxwell CSC (Promega) using the Maxwell DNA LEV Blood Kit (Promega #AS1290) according to manufacturer instructions. Libraries were prepared using the QIASeq Human Comprehensive Cancer Panel Kit (Qiagen #333,515) and sequenced on an Illumina HiSeq 2500 or NovaSeq 6000. All samples were collected in triplicate biological replicates at baseline and after achieving resistance to EGFR inhibition with persistent growth at 100% physiologic C_max. RNA libraries were constructed using TruSeq Stranded total RNA with rRNA reduction (Illumina). Quality control was performed by Eukaryote Total RNA electrophoresis using NanoDropOne. The library included 2 × 150 bp paired end reads with sequencing performed by NovaSeq 6000 using MiSeq NanoCell at the UW Biotechnology Center.

Subclonal analyses

Subclone counts were analyzed by counting the number of unique alterations present at a frequency between 10 and 30%, that were likely to alter the protein (missense mutations, insertions/deletions, and alterations to splice donors/ acceptors), and were not likely to be artifacts of sequencing. Artifacts in this case were defined as alterations that occurred in more than five individual patients. Subclones were analyzed using R version 3.6.1 (R Core Team 2019; https://www.R-project.org) and the tidyr v1.1.1⁶¹ and dplyr v0.8.5⁶² packages (https://CRAN.R-project.org). Relative variant allele frequency (rVAF) was defined by normalizing to the molecularly informed tumor content provided by the commercial vendor. Subclonal populations were reported in PCOs assuming 100% tumor content in the reads and defined with rVAF of 10–30%.

Variant calling and mutation analyses

Sequencing analysis through variant calling was performed at the UW Biotechnology Center. Sequence reads were adapted and quality trimmed using Skewer v0.2.2 (https://sourceforge.net/projects/skewer/)⁶³, aligned to Homo sapiens build 1k_v37 using BWA-MEM v0.7.17 (http://bio-bwa.sourceforge.net), and deduplicated using Picard v2.19 (http://picard.sourceforge.net) and Je v1.2 (https://anaconda.org/bioconda/je-suite)⁶⁴. Base quality scores were recalibrated using GATK v3.8 (https://gatk.broadinstitute.org/hc/en-us)⁶⁵ and mutations called using Strelka v-2.8.4 (https://github.com/Illumina/strelka)⁶⁶ without matched controls and annotated using SNPEff v4.3 (https://pcingola.github.io/SnpEff/)⁶⁷. Resulting VCF files were uploaded to the public Galaxy web platform⁶⁸ (usegalaxy.org) and cross-referenced to ClinVar’s publicly available VCF (accessed 2/13/2021) for annotation of predicted clinical response. Mutations were deemed pathogenic based on the ClinVar database labeled as either “Pathogenic” or “Likely Pathogenic.” Additionally, alterations in APC, KRAS, NRAS, BRAF, PIK3CA, EGFR, SMAD4, TP53, MLH1, MSH2, MSH6, PMS2, CTNNB1, and PTEN were evaluated for potentially pathogenic mutations not yet curated in ClinVar.

Statistical analysis of RNASeq data

All statistical calculations were performed using custom Rmarkdown scripts. The RSEM estimated counts were first filtered by removing Ensembl gene IDs that generated estimate counts of 0 across all samples then mapped to corresponding gene symbols. The pairwise (Resistance vs. Control for each cell line) differential gene expression was calculated using Bioconductor’s DESeq2 package v1.30 (DOI: https://doi.org/10.18129/B9.bioc.DESeq2)⁶⁹. For each sample, pathway expression values were calculated for the curated gene sets (c2) from MSigDB (https://www.gsea-msigdb.org/gsea/msigdb/) using GSVA v1.27 (DOI: https://doi.org/10.18129/B9.bioc.GSVA) and the Poisson model with the RSEM estimated counts⁷⁰, and KEGG pathway results are presented herein^71,72,73. Pair-wise differential pathway expression was calculated using limma v3.47 (DOI: https://doi.org/10.18129/B9.bioc.limma)⁷⁴. Using the aheatmap function from the NMF package v0.22 (https://cran.r-project.org/package=NMF)⁷⁵, heatmaps were made of the top 20 genes and top 20 pathways, scaling the regularized log2 (genes) or calculated enrichments across rows. Venn diagrams of the significant genes and pathways (p_adj < 0.05) were made using VennDiagram v1.6.22 (https://cran.r-project.org/web/packages/VennDiagram/index.html)⁷⁶.